A Neural Network embedded system for real-time identification of EMG signals

The objective of this work is to develop an embedded Artificial Neural Network (ANN) for the identification in real-time of the electromyography signal patterns (EMG). This system is going to be applied as an interface device between the user and a robotic prosthesis for upper limb. The methodology begins with the characterization of the extraction of the EMG signal: location of the sensors, sampling and definition of patterns; then the architecture and characteristics of the ANN are defined; then the ANN is implemented in an embedded system; and, finally, the cross-validation and validation in real time of the proposed system is carried out, by means of confusion matrices. The embedded ANN implemented is of the multilayer perceptron type, has 3 hidden layers, 27 neurons input layer, 4 neurons output layer, Fedforward architecture, sigmoid type transfer function, MSE error function and Backpropagation algorithm. In the evaluation of the system it was obtained that the average accuracy of the embedded ANN is higher than 97.7%, which confirms the reliability of using this type of systems as interface devices between a user and a robotic prosthesis. This development has a high practical implication, since an intelligent system is miniaturized, and can be part of any portable hardware device in an edge computing system.